Medical image segmentation method and system based on wavelet and visual mamba
By combining wavelet and visual Mamba, the scanning sequence of medical images is dynamically flattened, which solves the problem of the destruction of two-dimensional topological relationships in medical image segmentation by the visual Mamba model, and achieves high-precision segmentation of subtle lesions and integrity of anatomical structures.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHAANXI SCI TECH UNIV
- Filing Date
- 2026-04-17
- Publication Date
- 2026-06-26
Smart Images
Figure CN122289290A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of medical image segmentation technology, specifically to a medical image segmentation method and system based on wavelet and visual Mamba. Background Technology
[0002] Medical image segmentation, as a core supporting technology for computer-aided diagnosis and clinical decision-making, directly determines the reliability of lesion screening and surgical planning. In recent years, the visual Mamba model, based on a state-space model architecture, has become a research hotspot in the field of computer vision, achieving long-distance sequence dependency modeling while maintaining linear computational complexity.
[0003] However, unlike natural images, medical images typically possess strong structural continuity and weak boundary features, with lesion areas often exhibiting only subtle differences in grayscale or texture compared to surrounding normal tissue. Existing visual Mamba models, during feature extraction, typically employ a fixed spatial scanning order (such as row-by-row or column-by-column scanning) to rigidly flatten two-dimensional image blocks into a one-dimensional sequence. This fixed linear flattening method severely disrupts the inherent two-dimensional anatomical topology of medical images, causing previously spatially continuous high-frequency edges or minute lesion structures to be forcibly broken up in a one-dimensional sequence.
[0004] Because the true spatial adjacency relationship is destroyed by linearization, the model has difficulty effectively aggregating local detail features during state space transfer, resulting in a significant decrease in sensitivity to texture abrupt changes and small boundaries in the image, which in turn severely restricts the final segmentation accuracy of complex medical structures and lesion areas. Summary of the Invention
[0005] To address the aforementioned technical problems, the purpose of this application is to provide a medical image segmentation method and system based on wavelet and visual Mamba, and the specific technical solution adopted is as follows: In a first aspect, embodiments of this application provide a medical image segmentation method based on wavelet and visual Mamba, the method comprising the following steps: A medical image segmentation model based on wavelet and visual Mamba is constructed to obtain the feature map of the medical image to be processed from the input of the model; Multi-scale wavelet decomposition is performed on the feature map to be processed to obtain sub-band components of different frequency bands, and the high-frequency sub-band components are divided into multiple sub-maps. Based on the energy distribution characteristics of the subgraph within its subband component and the local structural change characteristics within the subgraph, the first index reflecting the importance of each subgraph under the high-frequency component is determined. A second index reflecting the structural continuity between adjacent subgraphs is determined by the consistency of changes in local background features of adjacent subgraphs under the same subband component. The first index is used to determine the starting subgraph for scanning, and the second index is used to determine the traversal path of the subsequent subgraphs, so as to determine the dynamic scanning order of the subgraphs in visual Mamba, and flatten the subgraphs into a one-dimensional sequence according to the dynamic scanning order. Feature extraction is performed on the one-dimensional sequence, and the sequence after feature extraction is restored to a two-dimensional feature map according to the spatial mapping relationship before flattening, which is used to obtain the segmentation result of the medical image input by the model.
[0006] In one embodiment, the medical image segmentation model employs the U-Net architecture.
[0007] In one embodiment, determining the first index includes: Calculate the sum of squares of the wavelet coefficients of all pixels in the sub-image as the energy value of the sub-image, and determine the proportion of the energy value of any sub-image in the total energy value of all sub-images in its sub-band component; Set a sliding window to slide on any of the sub-images and determine the range of wavelet coefficients for all pixels within each sliding window; The first index is calculated by combining the percentage with the numerical distribution of the range values corresponding to all sliding windows in any subgraph.
[0008] In one embodiment, the first index is calculated as follows: The mean of the range values corresponding to all sliding windows in any subgraph is calculated and positively fused with the percentage to obtain the first index.
[0009] In one embodiment, determining the second index includes: Determine the local spatial neighborhood of the target sub-image in its sub-band component, and extract the statistical distribution features of the feature parameters within the local spatial neighborhood as the local background features corresponding to the target sub-image; Calculate the feature deviations between the target sub-image and its neighboring sub-images in its local spatial neighborhood and the local background features respectively to obtain the first difference and the second difference; Calculate the similarity between the first difference and the second difference, and use the similarity as a second index that reflects the structural continuity between adjacent subgraphs.
[0010] In one embodiment, extracting the statistical distribution features of the feature parameters within the local spatial neighborhood as the local background features corresponding to the target sub-image includes: Calculate the mean wavelet coefficients of corresponding pixels in all sub-images within the local spatial neighborhood to obtain the average wavelet feature sub-image corresponding to the target sub-image; use the average wavelet feature sub-image as the local background feature corresponding to the target sub-image.
[0011] In one embodiment, determining the first difference and the second difference includes: Calculate the wavelet coefficient difference between the target sub-image and the corresponding pixel points in the average wavelet feature sub-image, and use the vector composed of all wavelet coefficient differences as the first difference; Accordingly, the wavelet coefficient differences between the corresponding pixel points in the average wavelet feature sub-image of the target sub-image and the neighboring sub-images in the local spatial neighborhood of the target sub-image are calculated to determine the second difference.
[0012] In one embodiment, determining the dynamic scan order of subgraphs in visual Mamba includes: Based on the first index of each subgraph, the subgraphs under the same subband component are divided into sets of subgraphs with different priorities, and the sets of subgraphs are traversed in order of priority from high to low. In the currently traversed set of subgraphs, the subgraph with the largest first exponent that has not been flattened is selected as the starting subgraph for the current scan. Based on the starting subgraph for the current scan, the unflattened subgraphs with the largest second exponent are selected sequentially as subsequent scan nodes until there are no unflattened subgraphs that meet the conditions in the starting subgraph for the current scan, thus forming a locally continuous scan sequence segment.
[0013] In one embodiment, if there are still unflattened subgraphs in the currently traversed set of subgraphs, the subgraph with the largest first index is selected as the new starting subgraph for scanning from the remaining unflattened subgraphs, and the step of selecting subsequent scanning nodes is repeated until all subgraphs in the set of subgraphs are flattened.
[0014] Secondly, embodiments of this application also provide a medical image segmentation system based on wavelet and visual Mamba, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.
[0015] This application has at least the following beneficial effects: Traditional Mamba scanning uses fixed row / column scans, which forcibly severs strong anatomical structures in medical images, resulting in the loss of spatial adjacency relationships in the sequence dimension. This application introduces wavelet decomposition to isolate high-frequency components representing details in the frequency domain. By fusing energy distribution and structural change features through a first exponent, the medical image segmentation model can automatically identify key sub-images containing potential lesions or edges, and use them as the starting point or high-priority region of the scan sequence. This avoids the dilution of key features in redundant background information, effectively improving the accuracy of capturing subtle and weakly boundary lesions. Furthermore, by quantifying the structural consistency between sub-images through a second exponent, the scan path can follow the natural extension of organ boundaries or lesion textures. This approach ensures that anatomically related features remain physically adjacent in the SSM (State Space Model) sequence, improving the integrity of structural edge segmentation and significantly enhancing the medical image segmentation model's ability to perceive long-range continuous structures. It achieves adaptive anatomical topology sequence modeling, overcoming the limitations of fixed scanning. This application can endow a one-dimensional sequence with two-dimensional spatial structural logic without increasing computational complexity, enabling the medical image segmentation model to retain the efficiency of Mamba in handling long-distance dependencies while compensating for the "structural breakage" or "void" phenomenon that traditional network models are prone to when processing fine anatomical structures in medical images. This allows the medical image segmentation model to take into account the consistency between the global receptive field and the local structure. Attached Figure Description
[0016] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 A flowchart illustrating the steps of a medical image segmentation method based on wavelet and visual Mamba provided in one embodiment of this application; Figure 2 This is a schematic diagram of a medical image segmentation model. Detailed Implementation
[0018] To further illustrate the technical means and effects adopted by this application to achieve the intended inventive objective, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the medical image segmentation method and system based on wavelet and visual Mamba proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.
[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0020] The following description, in conjunction with the accompanying drawings, details the specific scheme of the medical image segmentation method and system based on wavelet and visual Mamba provided in this application.
[0021] Please see Figure 1 The diagram illustrates a flowchart of a medical image segmentation method based on wavelet and visual Mamba according to an embodiment of this application. The method includes the following steps: First, a medical image segmentation model based on wavelet and visual Mamba is constructed. This model adopts a U-Net architecture that includes an encoder and a decoder. A schematic diagram of the medical image segmentation model is shown below. Figure 2 As shown, encoder 1 represents the first stage of the encoder, encoder 2 represents the second stage, encoder 3 represents the third stage, and encoder 4 represents the fourth stage. Similarly, decoder 1 represents the first stage, decoder 2 represents the second stage, decoder 3 represents the third stage, and decoder 4 represents the fourth stage. The encoder is responsible for extracting deep features from the input image through downsampling, while the decoder is responsible for upsampling the feature map layer by layer and combining skip connections to restore the original image size. Both the encoder and decoder are divided into four processing stages.
[0022] In the encoder's feature extraction process, a wavelet transform layer and a visual Mamba layer are cascaded after each downsampling stage. Specifically, the first downsampling stage of the encoder reduces the spatial size of the input medical image to 1 / 4 of its original size, and the subsequent three stages successively reduce the feature map size to 1 / 2 of the previous stage. After the downsampling operation in each stage, the current feature map is decomposed using two-dimensional discrete wavelet transform (2D-DWT) to obtain four sub-band components containing one low-frequency sub-band (LL) and three high-frequency sub-bands (horizontal LH, vertical HL, and diagonal HH), denoted as the wavelet feature map.
[0023] Furthermore, the wavelet feature maps of the four different frequency bands are input into the parallel visual Mamba layer for independent feature extraction. The specific processing flow of the visual Mamba layer is as follows: (1) Sub-map partitioning and coordinate recording: The wavelet feature maps of each frequency band are uniformly divided into multiple feature sub-maps of fixed size in the spatial dimension, and a spatial mapping table is established simultaneously to record the row and column coordinate indices (n, m) of each feature sub-map in the original two-dimensional feature map. (2) Dynamic sequence flattening: For the wavelet feature maps of the low-frequency sub-band (LL), the conventional fixed scanning order is used to flatten the sub-maps into a one-dimensional sequence. In this embodiment, row priority is used. For the wavelet feature maps of the high-frequency sub-bands (LH, HL, HH), the calculation method of this embodiment is used to determine the dynamic scanning order of each sub-map, and the two-dimensional feature sub-maps are flattened into a one-dimensional sequence according to the dynamic scanning order. (3) State space update: The state space model (SSM) inside the visual Mamba is used to perform global long-distance dependency modeling and feature update on the flattened one-dimensional sequences of each frequency band. (4) Inverse mapping restoration and multi-band fusion: For the one-dimensional sequence after the state space is updated, since the dynamic scanning breaks the continuity of physical storage, according to the spatial mapping table recorded in step (1), each feature element in the one-dimensional sequence is inversely mapped back to the corresponding position of the two-dimensional matrix according to its original coordinate index (n, m), and restored to a two-dimensional feature map of four independent frequency bands; then, the four restored two-dimensional feature maps are spliced in the channel dimension to form a fused feature map, and the channel number is compressed and the features are fused through a 1×1 convolutional layer to obtain the new feature map of the final output of this stage.
[0024] The fused feature maps are fed into the decoder of the U-Net architecture. The upsampling layer of the decoder is also divided into four stages, progressively amplifying the new feature maps. Through step-by-step upsampling and skip connections, shallow spatial details are supplemented, ultimately restoring the image to its original size to obtain the segmentation result of the input medical image. The optimizer used in training the medical image segmentation model is the Adam optimizer, and the loss function is the Dice loss function.
[0025] In medical image segmentation models, the process of determining the dynamic scanning order of each sub-image is as follows: Step 1: Obtain the feature map of the medical image input to the medical image segmentation model. Perform multi-scale wavelet decomposition on the feature map to obtain sub-band components of different frequency bands, and divide the high-frequency sub-band components into multiple sub-maps. Specifically: Traditional visual Mamba images segment feature maps into fixed-size sub-images and unfold them into a one-dimensional sequence according to a preset scanning order, such as row-first or column-first, for state-space modeling computation. However, medical images are characterized by strong structural continuity and weak boundaries, and organ structures have clear spatial topology and anatomical coherence. Unfolding images in the traditional scanning order disrupts the original two-dimensional spatial adjacency relationships, causing spatially adjacent and structurally related regions to be scattered in the sequence, while spatially discontinuous regions may be arranged adjacently, resulting in a weakening of structural information during state propagation. This reduces the model's sensitivity to fine-grained lesion boundaries and subtle structural changes, thus affecting its ability to accurately represent complex medical structures and ultimately leading to poor accuracy and stability in the segmentation of small lesions.
[0026] Therefore, based on the above analysis, the feature map obtained after each downsampling stage in the encoder is denoted as the feature map to be processed, and the feature map to be processed in the first stage is taken as an example. For example, the feature map to be processed Multi-scale discrete wavelet decomposition is performed using wavelet basis functions (such as Haar, Daubechies, etc., without specific restrictions). Considering that the image is downsampled at each stage, resulting in a decrease in image size, the wavelet decomposition layer is set to 1 in this embodiment, yielding wavelet feature maps in four different frequency bands (LL, LH, HL, HH). The wavelet basis function used in this embodiment is the Haar function; implementers can choose other wavelet basis functions. Discrete wavelet decomposition of images is a well-known existing technique and will not be elaborated further.
[0027] Furthermore, wavelet feature maps of different frequency bands are input into the visual Mamba layer. First, visual Mamba segments each wavelet feature map, resulting in the feature map to be processed. The number of sub-maps segmented by the wavelet feature map is indivual, It is a multiple of 2, and its size can be set by the implementer according to the size of the feature map in the implementation scenario. In this embodiment... Then the sub-images after segmentation in the wavelet feature map. The spatial mapping table is denoted as , This indicates that the spatial position of the subgraph in the wavelet feature map is the th. Line 1 This column records the spatial mapping table of the subgraph, which is used to restore the spatial correspondence in the feature map in the subsequent unfolded sequence.
[0028] Step 2: Based on the energy distribution characteristics of the subgraph within its respective subband component and the local structural change characteristics within the subgraph, determine the first index that reflects the importance of each subgraph under the high-frequency component.
[0029] In the wavelet feature map obtained by multi-scale wavelet decomposition, the LL band, as a low-frequency sub-band, is mainly used to characterize the macroscopic contours of organs and the continuous distribution features of global tissue structures in medical images; the LH, HL and HH bands, as high-frequency sub-bands, are used to capture local high-frequency detail information of abrupt changes in gray level and structural abrupt changes in the horizontal, vertical and diagonal directions of the image, respectively.
[0030] In clinical medical images, lesions, compared to normal tissue, often exhibit disruption of the original structural continuity, accompanied by disordered local texture features and dramatic fluctuations in grayscale gradients. Therefore, the feature information rich in the high-frequency subband closely matches the apparent morphology of the lesion region, accurately mapping potential lesion boundaries and small, important areas in the image, providing a reliable basis for subsequent extraction of key structural features.
[0031] Therefore, a first index is constructed by analyzing the energy statistical characteristics and structural change characteristics of subgraphs under different high-frequency subbands (LH, HL, HH) to reflect the importance of a single subgraph in the entire wavelet feature map.
[0032] Specifically, taking the wavelet feature map under the LH subband as an example, the submap energy value This is expressed as the sum of squares of the wavelet coefficients of all pixels in the sub-image. Then, the cumulative sum of the energy values of all sub-images under the LH sub-band is calculated. Then the subgraph The first parameter The calculation method is as follows If the sum of the energy values of all subgraphs under the LH subband is 0, then all subgraphs of the wavelet feature map under the LH subband are flattened into a one-dimensional sequence using a conventional fixed scanning order.
[0033] The first parameter is the relative proportion of the energy of a single sub-image in the entire high-frequency sub-band, which reflects the degree of drastic change in the global structure of the sub-image. If the energy proportion of the sub-image is high, it indicates that there are obvious structural or local texture mutations within it, which may correspond to organ boundaries or lesion features, and to a certain extent reflects the importance of the sub-image in the entire frequency band.
[0034] Furthermore, using a side length of A sliding window of 1 pixel in the subimage Slide within the subgraph and perform statistical analysis. The structural change characteristics within. Among them, It should be a multiple of 2; if the wavelet feature map size is large, it can be increased accordingly. The size, in this embodiment Using a sliding window The process involves sliding the wavelet coefficients within a single sliding window, and each slide counts the range of the wavelet coefficients for all pixels in that window. The magnitude of the range reflects the difference and unevenness of structural changes in the local pixel region of the subimage. A large range indicates a significant abrupt change in grayscale or an edge transition within the sliding window; while a small range indicates a relatively smooth local structural change and a relatively uniform and stable region.
[0035] Furthermore, statistical subgraphs The average of the ranges of wavelet coefficients across all sliding windows is used as the second parameter. Then, the second parameter of all subgraphs within the same wavelet feature map is calculated, and the second parameter of all subgraphs is normalized using a maximum-minimum method. The normalized second parameter is It should be noted that when performing maximum-minimum normalization, if the denominator is 0, the second parameter of all subgraphs in the same wavelet feature map will be proportionally reduced to the [0,1] interval.
[0036] For the first parameter With the second parameter Perform forward fusion to obtain a subgraph The first index. Here, positive fusion represents combining multiple variables, which can be calculated using methods such as addition, multiplication, or weighted summation.
[0037] In this embodiment, subgraph The expression is: In the formula, , The weight coefficients of the first parameter and the normalized second parameter are respectively, satisfying... In this embodiment , Take values of 0.5 and 0.5 respectively.
[0038] When the first parameter and the normalized second parameter are large, it indicates that there are obvious directional structural changes or local texture mutations in the sub-image, and the magnitude of the structural changes varies greatly. This may correspond to lesions or organ edges. In this case, the importance of the sub-image is relatively high compared to structurally stable organ regions.
[0039] Step 3: Determine the second index that reflects the structural continuity between adjacent sub-images by the consistency of changes in local background features of adjacent sub-images under the same sub-band component.
[0040] While a single sub-image may reveal grayscale and texture variations within a very small spatial range, organ boundaries and lesions in medical images typically exhibit continuous distribution and spatial correlation. Local sub-images have limited ability to express long-range anatomical relationships in medical images. Conventional sub-image flattening strategies lead to the linear decomposition of spatially adjacent regions with strong structural coupling characteristics after conversion into a one-dimensional sequence. This weakens the originally continuous lesion edges or organ contour information during sequence propagation, making it difficult for the model to accurately identify subtle structural evolutions and affecting the final segmentation stability.
[0041] Based on the above description, this embodiment constructs a second index by analyzing the structural change characteristics of subgraphs and adjacent subgraphs within the local neighborhood under different high-frequency subbands (LH, HL, HH), which reflects the structural continuity between different subgraphs.
[0042] Specifically, taking the wavelet feature map under the LH subband as an example, for the submap... subgraph The subgraph within the eight neighborhoods of the center is used as the subgraph Local neighborhood range Let be denoted as the local spatial neighborhood; the range of this local neighborhood includes the subgraph. The mean of wavelet coefficients of corresponding pixels in all sub-images is used to obtain the local neighborhood range. Average wavelet feature subgraph of all subgraphs as a subgraph The local background features. First, calculate the subgraph. and The wavelet coefficient differences between corresponding pixels are calculated, and all the calculated wavelet coefficient differences are arranged into a vector according to the positional order of the pixels, denoted as the first difference. Then, calculate the range belonging to the local neighborhood. Subgraph within and The wavelet coefficient differences between corresponding pixels in the sub-image The differences of all wavelet coefficients calculated are arranged into a vector according to the positional order of the pixels, and denoted as the second difference. The size of the local neighborhood can be set by the implementer according to the actual situation; this embodiment does not impose any restrictions on this.
[0043] It should be noted that, since the sub-images are all the same size, the corresponding pixel in this embodiment refers to the pixel in the same position within the sub-image. For example, if the pixels in the sub-image are arranged from left to right and from top to bottom, then the sub-image... The first pixel in The first pixel in the average wavelet feature sub-map is the pixel at the corresponding position. In addition, the difference represents the degree of difference between two variables. Specifically, it can be calculated by methods such as the absolute value of the difference and the square of the difference. In this embodiment, the wavelet coefficient difference is calculated by the absolute value of the difference.
[0044] Calculate the similarity between the first difference and the second difference, and use the similarity as the second index reflecting the structural continuity between adjacent sub-maps. In this embodiment, the sub-map and its adjacent sub-maps within its local neighborhood range The second index between Specifically, it is the first difference and the second difference The cosine similarity between them.
[0045] The first difference and the second difference reflect the relative structural change characteristics of the sub-map in the current local neighborhood space range relative to the common local structural background. In this embodiment, the second index represents the degree of consistency between the sub-map and the sub-map in the direction of relative structural change, that is, whether their change trends relative to the common local structural background are similar, thereby reflecting the structural continuity and extensibility in the local space. If the second index is large, it means that the sub-map and the sub-map belong to the same continuous structure or edge extension in the local neighborhood; on the contrary, the smaller the second index value, the more obvious the break in the structure between the sub-map and the sub-map , and the worse the structural continuity.
[0046] Step 4: Determine the starting scanned sub-map based on the first index, and determine the traversal path of the subsequent sub-maps based on the second index to determine the dynamic scanning order of the sub-maps in the visual Mamba, and flatten the sub-maps into a one-dimensional sequence according to the dynamic scanning order.
[0047] Taking the wavelet feature map of the LH sub-band as an example, the scanning order for expanding each sub-map in the wavelet feature map of the LH sub-band into a sequence is defined as follows. Specifically: According to the numerical values of the first indices of each sub-map, the sub-maps under the LH sub-band are divided into sub-map sets with different priorities. In this embodiment, the value range of the first index is divided into three equal intervals. Among them, the sub-maps with the first index are classified into the set with the highest priority, and the sub-maps with the first index are classified into the set Second in priority, first index Subgraphs are grouped into sets It has the lowest priority.
[0048] First, in the subgraph set Select the subgraph with the largest first index. As the first target to be developed Then, in the subgraph set Further search for subgraphs Subgraph with the largest second exponent As the second expansion target Then, continue in the subgraph set Finding subgraphs in unflattened subgraphs The subgraph with the largest second exponent is used as the third expansion object. until the subgraph set Middle and Subgraph After all subgraphs with a second index that is non-zero are expanded, determine the subgraph set. Does a subgraph still exist? If so, select the subgraph with the largest first index. Continue expanding into a sequence, repeating the above steps to select and subgraphs. The subgraph with the largest second exponent is selected as the subgraph. The subsequent expansion of objects continues until the subgraph set is included. All subgraphs in the set can be unfolded into a one-dimensional sequence. It should be noted that if the subgraph set... If there are still unexpanded subgraphs, but all of them have the same or zero first index, then the subgraph corresponding to the minimum row coordinate is selected as the subgraph with the largest first index. From the remaining unexpanded subgraphs, the subgraph with the largest second index is then selected sequentially and expanded until all unexpanded subgraphs are expanded. Specifically, if there are multiple subgraphs corresponding to the minimum row coordinate, then the subgraph corresponding to the minimum y-coordinate is selected as the subgraph with the largest first index.
[0049] Correspondingly, for the set of subgraphs Subgraph set All subgraphs are sequentially expanded into a one-dimensional sequence. Following the expansion method of each subgraph of the wavelet feature map of the LH subband, each subgraph of the wavelet feature map of the HL and HH subbands is expanded to complete the scanning order strategy for sequence flattening of visual Mamba.
[0050] Based on the same inventive concept as the above methods, this application also provides a medical image segmentation system based on wavelet and visual Mamba, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above-described medical image segmentation methods based on wavelet and visual Mamba.
[0051] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments of this specification have been described above. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.
[0052] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
[0053] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the principles of this application should be included within the protection scope of this application.
Claims
1. A method for medical image segmentation based on wavelet and visual Mamba, characterized in that, The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps:
2. The wavelet and visual Mamba based medical image segmentation method as claimed in claim 1, wherein, The method comprises the following steps:
3. The wavelet and visual Mamba based medical image segmentation method as claimed in claim 1, wherein, The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps:
4. The wavelet and visual Mamba based medical image segmentation method as claimed in claim 3, wherein, The method comprises the following steps: The method comprises the following steps:
5. The wavelet and visual Mamba based medical image segmentation method as claimed in claim 1, wherein, The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps:
6. The wavelet and visual Mamba based medical image segmentation method as claimed in claim 5, wherein, The method comprises the following steps: The method comprises the following steps:
7. The wavelet and visual Mamba based medical image segmentation method as claimed in claim 6, wherein, The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: Correspondingly, a wavelet coefficient difference between pixels at corresponding positions in adjacent subgraphs within a local spatial neighborhood of the target subgraph and in average wavelet feature subgraphs corresponding to the target subgraph is calculated, and a second difference is determined.
8. The wavelet and visual Mamba based medical image segmentation method as claimed in claim 1, wherein, The method further includes determining a dynamic scanning order of the subgraphs in the visual Mamba, including: According to the first index of each subgraph, subgraphs under the same subband component are divided into subgraph sets of different priorities, and the subgraph sets are traversed in order from high to low priority; In the currently traversed subgraph set, a subgraph with the largest first index and not flattened is selected as a current scanning starting subgraph; based on the current scanning starting subgraph, an unflattened subgraph with the largest second index from the current scanning starting subgraph is selected as a subsequent scanning node, until there is no unflattened subgraph that meets the condition for the current scanning starting subgraph, forming a locally continuous scanning sequence segment.
9. The wavelet and visual Mamba based medical image segmentation method as claimed in claim 8, wherein, If there are still unflattened subgraphs in the currently traversed subgraph set, the subgraph with the largest first index among the remaining unflattened subgraphs is selected as a new scanning starting subgraph, and the step of selecting a subsequent scanning node is repeated until all subgraphs in the subgraph set are flattened.
10. A medical image segmentation system based on wavelets and visual Mamba, comprising a memory, a processor and a computer program stored in the memory and running on the processor, characterized in that, The processor executes the computer program to implement the steps of the method of any one of claims 1-9.